Ivx Content Factory Pipeline

Content Factory pipeline runner

Content Factory is the intelli-verse-x media generation backend. It exposes ~50 named pipelines (each backed by a YAML config in configs/pipelines/) via a FastMCP server on port 8001.

When to use this skill

• User asks for video / image / audio / screenshot generation by name (learning_series, movie, kids_movie, ad, ad_banners, quiz_shorts, podcast2video, documentary, app_store_deployer, game_marketing_audit, song, dubbing, world_scene, …).
• User asks to "make content for X app" or "produce a trailer for Y".
• Status checks on a previously-triggered run (get_task_status).
• Listing what pipelines are available.

Prerequisites

The MCP servers from this repo are configured in the calling client:

• content-factory — FastMCP at mcp_server/server.py (HTTP-proxies CF api)
• content-factory-media — FastMCP at media_mcp_server/server.py (provider-agnostic generate_image/video/tts/music/motion)

Env vars the MCP server reads: CF_API_URL (default http://localhost:8001), CF_API_KEY, DIRECTOR_URL (default http://localhost:8000).

The workflow

1. Pick the pipeline. If the user named one, use it. If not, ask for one of: learning_series, movie, kids_movie, short_movie, short_video, video_shorts, quiz_shorts, podcast_series, podcast2video, documentary, blog, ebook, song, song2musicvideo, beat_synced, dubbing, ad, ad_banners, app_ad_campaigns, event_promo, event_recap, marketing_kit, gtm_master_plan, revenue_strategy, game_visuals, game_sound, ivx_full_game, ivx_character_2d, ivx_character_3d, ivx_landing_page, world_scene, character_2d, character_3d, motion_library, interactive_avatar, comic, lyrics, legal_gen, screenshot_localizer, asc_release, aso_keywords, app_catalog_enricher, app_store_deployer, competitor_analysis, content_planner, guided_learning, short_drama, short_movie_series, long_form_video, script2video, video, video_shorts.

2. Build the request body. Every pipeline accepts a config_path (or inline config), an output_dir, and a brand_context / game_context block. Defaults in configs/pipelines/<name>.yaml.

3. Call trigger_<pipeline_name> (or generic trigger_pipeline) on the content-factory MCP server. It returns a task_id.

4. Poll get_task_status(task_id) every 30s. Status flows: queued → running → done | failed | blocked. If blocked, read the comment thread — usually a missing asset or a quality gate failure.

5. Surface artefacts. On done, the response includes:

   • output_dir (S3 or local path)
   • artefacts[] — list of (kind, uri) tuples
   • quality_score if the pipeline has a quality.py gate
   • council_log for any council-voted creative outputs

Examples

Learning series episode

trigger_learning_series(
  config_path="configs/pipelines/learning_series.yaml",
  topic="Newton's three laws of motion",
  episode_count=3,
  locale="en",
  brand_context={"brand": "QuizVerse", "audience": "ages 10-14"}
)

Game trailer (one-shot)

trigger_ivx_full_game(
  config_path="configs/pipelines/ivx_full_game.yaml",
  game_name="Cricket VR",
  game_context={"genre": "sports", "platform": "iOS, Android, Quest"}
)

App Store screenshots, all locales

trigger_screenshot_localizer(
  config_path="configs/pipelines/screenshot_localizer.yaml",
  app_slug="quizverse",
  locales=["en", "es", "fr", "de", "ja", "ko", "zh-Hans", "pt-BR"]
)

Quality + cost guardrails

• Always confirm brand_context before triggering — pipelines that miss it produce off-brand artefacts and burn credits.
• For long-running pipelines (learning_series with episode_count > 5, any movie pipeline), prefer queuing through Hermes Kanban instead of polling — see the ivx-game-trailer skill for the pattern.
• Surface the council_log to the user when present — it shows which model voted for what and is the audit trail for creative choices.

Common failures